Multilingual TTS Model — 40+ Languages, 100+ Dialects, <300ms Latency
A production multilingual TTS with voice cloning — 40+ languages, 100+ dialects, and <300ms latency. ElevenLabs-class quality, self-hosted, trained on 8×H100 GPUs.
What Voice-first SaaS client was up against
ElevenLabs-class TTS was commercially perfect for the client's product — but the price per character and vendor dependency made it unworkable at their target scale. Off-the-shelf open TTS had either insufficient voice quality, narrow language coverage, or no practical way to clone a new voice from a short sample. They wanted ElevenLabs-level quality across 40+ languages and 100+ regional dialects, with sub-300ms time-to-first-audio, all running inside their own infrastructure.
What we built
We started from open TTS architectures (XTTS-style, VITS-derived), curated and cleaned a multilingual training corpus spanning 40+ languages and 100+ dialects, and fine-tuned the model on an 8× H100 GPU cluster with prosody-preserving augmentation. Voice cloning works from a 30-second reference sample. The model ships with a streaming inference server optimized for real-time use, a language-expansion pipeline so the client can onboard new dialects without our help, and built-in safety filters plus voice-consent gating.
What shipped
Want something similar?
Other work we’ve shipped
Telluswhatyouwanttoautomate.We'llreplyinonebusinessday.
Describe the problem, the constraint, the deadline. We'll send back a scoped plan and a senior engineer to kick it off — no sales theater.